The brief was a folder-based AI coach. What we built is a formed mentor — and the distinction is not semantic.
A coach is a service. You pay for access to a framework. When a coach pushes back, you can dismiss it — you hired him. His opinion is a product. A mentor chooses you as much as you choose them. When a mentor pushes back, it lands differently. He didn't have to choose you. This governing decision changed everything about the architecture: the character layer is fixed, not reactive. The Praeceptor is not a detection system that scans input and deploys matching content. He arrives with something — a point of view built from the composite of Andy Grove's ruthless clarity, Charlie Munger's oblique wisdom, Bill Campbell's courage evangelism, Bill Walsh's standard-setting patience, George Marshall's selfless authority, Seneca's urgency about time, Marcus Aurelius's daily self-interrogation, and Naval Ravikant's first-principles clarity. He has carried failures that belong to him, not to a profile analysis of the person across from him. Every session starts with the same CHARACTER layer — immutable, fully loaded — that cannot be compressed or crowded out by context. The KNOWING layer (≤800 tokens, hard limit) tells him who he's talking to. It informs his judgment. It does not change who he is.
The research behind the character is sourced, not approximated. Ninety-four thousand words of transcript across thirteen sources — both Munger DJCO sessions and the Psychology of Misjudgment, the full Robert Greene interview, the Acquired Intel episode, Grove at Stanford, Walsh on the Standard of Performance, Campbell's 33 principles, Kim Scott on radical candor. The three books whose authors define the character most directly (Seneca, Aurelius, Naval) were fetched from public domain and processed. The remaining nine books were extracted from the firm's research library. A NotebookLM notebook with all 25 sources produced seven targeted research outputs that fed directly into the ICM files: specific quotes from each composite figure with context, named frameworks with behavioral rules, failure stories with deployment instructions, voice register analysis for each source figure, the non-negotiable refusals, and seeing-clearly practices from Grove's variable inspection to Aurelius's morning preparation.
The iOS application is not a demo wrapper. It is a production-grade voice interface built in Swift 6 with SwiftUI targeting iOS 18. The voice pipeline: hold to speak → AVFoundation recording → release → OpenAI Whisper transcription → assembled system prompt (CHARACTER + KNOWING + session history) → Claude Sonnet 4.6 streaming via AsyncThrowingStream → OpenAI TTS (voice: onyx, speed: 0.92) → AVFoundation playback. Three time-of-day visual modes change not just the interface but the system prompt modifier and TTS speed: morning is amber/gold and forward-looking at 0.92, noon is grey/blue and execution-focused at 0.95, night is navy/earth and reflective at 0.88. The intake flow ingests available context first, identifies gaps, then asks a maximum of seven dynamic questions — only what's actually missing, never formula questions. The KNOWING layer persists to iCloud between sessions. The mentor reads it once at session open. He doesn't check his notes while talking to you.